Pigeons and AI have more in common than you might think
David Lock, senior AI product manager
I am a great fan of the ‘No such thing as a fish’ podcast, a weekly broadcast of trivial yet hugely interesting facts and discussion from the makers of the QI panel show from the BBC.
During one a recent podcast, the team highlighted a case a few years ago in which scientists were able to train pigeons to spot breast cancer from MRI scans, thanks to their excellent eyesight, with an individual pigeon success rate score of 73%.
Interestingly, if the images were ‘flock-sourced’ e.g. the same image was shown to multiple birds, results went up to 99% accuracy in spotting the tumours.
We have used AI as part of Reno-Secure to replicate this ‘flock-sourcing’ approach to problem-solving, with equally impressive results.
Using multiple machine learning algorithms to simulate different pigeons’ ‘opinions’, we are able to simulate ‘flock-sourcing’. Different algorithms with different capabilities and varied parameters can be run simultaneously to provide a range of values. What we also do, using the same data, is run different algorithms to detect multiple problems simultaneously.
Our heuristics engine then allows rules and multiple machine learning models to be combined to provide a single desired target. For example, in the context of an ATM, we can ask: does this ATM need maintenance to prevent an impending failure? The simulated ‘flock-sourcing’ allows us to provide more accurate results and reach an informed decision than otherwise would have been possible.
Given today’s computing power and highly available data, a single machine learning model is not the answer. The ability to quickly and efficiently manage and combine models in ensembles run in a cloud environment is the key to accurate and precise automated fleet management.
And then there is, of course, the added advantage of not having to clear up after the pigeons!